Claude Skills by mangit955

Use when adding a new tool to tools/, or changing an existing tool's name, parameters, description, or output shape. Covers the registry, the effects table, the generated docs surface, and the invariants a tool must hold (truncation, workspace containment, approval, AbortSignal, model-facing errors). Triggers on "add a tool", "new tool", "the agent should be able to ...", or any edit under tools/.

Use when changing how a shell command is judged or gated — anything under runtime/approval/, or the classification tables in runtime/toolEffects.ts. Fail-closed is the invariant, every change needs a fixture in both directions, and both operating systems' path semantics have to be reasoned about. Triggers on "approval mode", "auto-approve", "this command shouldn't need confirmation", "the agent keeps asking", or any edit to the classifier or the effects table.

Use when writing or changing any test file in this repository. Covers the module-mock trap that once made the whole suite green for the wrong reason, testing against real files instead of mocks, the environment a test runs in, and the sweep that proves the suite is honest. Triggers on any edit to a *.test.ts or *.test.tsx file, "add a test", "write a regression test", or a failing suite that passes in isolation.

Use when changing anything that decides what goes into the prompt — runtime/loop.ts, runtime/compaction.ts, runtime/executionLog.ts, config/systemPrompt.ts, or buildRepositoryContext and recentMessages in config/config.ts. Requires a replay baseline captured before the edit and compared after, because the harness that measures this is opt-in and the gate does not know it exists. Triggers on "context window", "compaction", "token budget", "system prompt", "the agent is sending too much", or an...